ISCA Archive SPECOM 2004
ISCA Archive SPECOM 2004

Double clustering algorithm applied to speaker dependent information

J. Srinonchat, S. Danaher, J. I. H. Allen, A. Murray

The clustering algorithms are used to group or cluster similar input vectors together. This technique has been widely applied to linear and non-linear system for propose of classifying signal. In this paper, clustering algorithm has been work for speech signal. K-means and Self Organising Maps (SOM) methods are very useful and popular as performance as clustering techniques. Due to SOM has a property of self-organising that all nodes in the network are connected and updated all nodes with a particular neighbourhood function. In this reason, SOM is selected to perform under the condition of clustering on information of speech data within Speaker Dependent system. Speech information in Speaker dependent system exists both in the spectral envelope and in the voice source characteristics of speech. The spectral envelope is the characteristics of the vocal tract and the voice source characteristics acquires from the different in manner of speaking such as dialect and accent. This individual information can be feature classifies and exploits in many of speech research works. However to accomplishment in coding, synthesis or recognition speech signal of a particular single speaker, it is necessary to understand the characteristic of that speech. This paper is introduced a novel model of clustering technique, namely Double Clustering Self-Organising Maps which based on Self-Organising Maps, to classify and cluster the information of speaker dependent and also encourage this technique to further compression speech signal. This techniques is difference from the traditional SOM, in this work, the network structure in DCSOM is adaptively adjusted itself during the learning phase, so that the similar characteristic neural will have similar weight vectors and also move nearer to each other. Then the group of neuron which has similar property is clustered using the specific distance property function. The result shows that by this technique, it can generate and determine the optimum number of clustering in an efficient and effective way for speech signal especially on speaker dependent system. The main result of the new model that uses this technique shows the effective performance of the accuracy classify in each speaker dependent and also can reduce bit rate of speech signal about 33 % in the environment of speaker dependent coding system which can be applied to Computer Base Learning (CBL).

Cite as: Srinonchat, J., Danaher, S., Allen, J.I.H., Murray, A. (2004) Double clustering algorithm applied to speaker dependent information. Proc. 9th Conference on Speech and Computer (SPECOM 2004), 371-376

  author={J. Srinonchat and S. Danaher and J. I. H. Allen and A. Murray},
  title={{Double clustering algorithm applied to speaker dependent information}},
  booktitle={Proc. 9th Conference on Speech and Computer (SPECOM 2004)},